MobileNetV2 风格的 Inverted Residual block。

MobileOne building block. This block has a multi-branched architecture at train-time and plain-CNN style architecture at inference t

Multi-Head Attention layer. Args: n_head (int): The number of heads. n_feat (int): The number of features. dropout_rate (

音频驱动的轻量数字人 UNet。 输入： - x: [B, n_channels=6, 160, 160] (BGR ref + masked current) - audio_feat: [B, 128, 16, 32] (wenet) 或 [B, 1

Positionwise feed forward layer. FeedForward are appied on each position of the sequence. The output dim is same with the input dim. Arg

Base class of Transfomer decoder module. Args: vocab_size: output dim encoder_output_size: dimension of attention attentio

Base class of Transfomer decoder module. Args: vocab_size: output dim encoder_output_size: dimension of attention attentio

CTC module

Encoder layer module. Args: size (int): Input dimension. self_attn (torch.nn.Module): Self-attention module instance.

Single decoder layer module. Args: size (int): Input dimension. self_attn (torch.nn.Module): Self-attention module instance.

Label-smoothing loss. In a standard CE loss, the label's data distribution is: [0,1,2] -> [ [1.0, 0.0, 0.0], [0.0, 1.0, 0

Positional encoding. :param int d_model: embedding dim :param float dropout_rate: dropout rate :param int max_len: maximum input length

Encoder layer module. Args: size (int): Input dimension. self_attn (torch.nn.Module): Self-attention module instance.

The WarmupLR scheduler This scheduler is almost same as NoamLR Scheduler except for following difference: NoamLR: lr = optimizer

CTC-attention hybrid Encoder-Decoder model

CTC-attention hybrid Encoder-Decoder model

Base class of Transfomer decoder module. Args: vocab_size: output dim encoder_output_size: dimension of attention attentio

Base class of Transfomer decoder module. Args: vocab_size: output dim encoder_output_size: dimension of attention attentio

Conformer encoder module.

Conformer encoder module.

VGG19 conv3_3 特征上的 MSE 感知损失。VGG 仅作特征提取器，禁用梯度。

Squeeze and Excite module. Pytorch implementation of `Squeeze-and-Excitation Networks` - https://arxiv.org/pdf/1709.01507.pdf

Transformer encoder module.

Transformer encoder module.

0,1,...,N-2,N-1,N-2,...,1,0,1,... 来回索引，每次 advance() 自动转向。

按照 0,1,2,...,N-1,N-2,...,1,0,1,... 的来回顺序无限取帧。

hubert 输入：[B, 16, 32, 32] → [B, 512, H', W']

wenet 输入：[B, 128, 16, 32] → [B, 512, H', W']

Convolutional 2D subsampling (to 1/4 length). Args: idim (int): Input dimension. odim (int): Output dimension. dropout_ra

Convolutional 2D subsampling (to 1/6 length). Args: idim (int): Input dimension. odim (int): Output dimension. dropout_rat

Convolutional 2D subsampling (to 1/8 length). Args: idim (int): Input dimension. odim (int): Output dimension. dropout_ra

ConvolutionModule in Conformer model.

Linear transform the input without subsampling Args: idim (int): Input dimension. odim (int): Output dimension. dropout_r

No position encoding

Multi-Head Attention layer with relative position encoding. Paper: https://arxiv.org/abs/1901.02860 Args: n_head (int): The number of

Relative positional encoding module. See : Appendix B in https://arxiv.org/abs/1901.02860 Args: d_model (int): Embedding dimension.

Construct an Swish object.